[Translate to Englisch:]
The Canadian computer scientist Yoshua Bengio and the mathematician and physicist Fabian Theis are now working together with their teams in the Helmholtz International Lab "Causal Cell Dynamics" (Image: collage from Maryse Boyce; Technische Universität München TUM/Astrid Eckert).

“We can’t wait to combine our expertise”

The Helmholtz International Lab "Causal Cell Dynamics" is now collaborating with the AI research institute Mila from Canada. The goal: to understand cellular processes in the development of diseases with the help of Big Data and AI. Fabian Theis and Yoshua Bengio explain the idea. 

Yoshua, you are the founder and director of the Mila – Quebec AI Institute with approximately 500 researchers in the field of AI. What is it that Fabian and his team can offer that you don’t have already?

Yoshua Bengio: We are one of the biggest AI research institutions in the world. We conduct research using artificial intelligence in many different fields where machines or humans can learn from data or from interacting with their environment, and we are best known for our work in deep learning and reinforcement learning. Machine learning is relevant in many areas, for example biomedical research, robotics, natural language processing, cognitive neuroscience, and computer vision. One major focus is of course on the design and analysis of algorithms, and the development of responsible AI applications, i.e., for the benefit of humanity. Fabian and his team are specialized and experienced in cellular dynamics, which raises fundamental AI research questions around how to model cause and effect, one of the hottest topics in machine learning these days. What makes it even more interesting for us is the approach taken by the Institute of Computational Biology team led by Fabian, focusing on single cell technologies.

What is the idea behind single cell technologies?

Fabian Theis: A few years ago, the enormous potential of AI for understanding cellular processes might have sounded far-fetched. But the technological progress made recently in single cell genomics, i.e. the collection of cellular state data over the last few years, is incredible. In traditional genomics, we usually analyze the average across a big number of cells, often 100,000 or more. As a result, you get the mean development of many single cells, which often obscures actual cellular processes. Therefore, what most experiments look at is actually a big mix, like a smoothie. In contrast, with single cell genomics we observe what happens in single cells – we look at the fruit salad, so to speak, before it is blended together as a smoothie. Seeing this level of detail can provide deeper insights.

But is what one single cell does really important? 

Fabian Theis: Cells are what we are built of. The communication and metabolism of cells is what sums up human life. Thus, it plays an elementary role in most diseases. For example, diabetes is the result of a change in cellular processes. Even though we look at single cells, we do not, of course, only look at one single cell but rather at many single cells, millions even, with more than 20,000 measurements per cell; true big data. Analyzing that data using AI methods, we have already identified many promising structures and critical mechanisms to predict the further development of the cells. We want to take the next step with the AI experts from Mila. 

Yoshua Bengio: We know each other from research visits; it was in the air that we would collaborate one day. Now that it’s getting started, we are excited: we can’t wait to work with the large amounts of biodata from our Helmholtz colleagues and to combine the experience of our research groups to further analyze, compare, structure and search these data sets. The goal is to develop machine learning methods to capture cell dynamics, i.e., to help us better understand how cells act and react. Because the more we know about the underlying algorithms and mechanisms, the better we understand why cells change in certain diseases – and what might usefully intervene in these processes… 

Fabian Theis: What is important here is that we are not simply searching for correlations. We want to go one step further now. Among the correlations, we want to find the ones that are relevant to health and disease. We want to understand – and also to predict, using our models – the significant causalities: What effect exactly does a certain drug have on the metabolism of single cells? What changes or defects can cause the development of certain diseases? And how do certain effects interact with which outcomes?

Do you have a specific example of this causal approach being applied?

Yoshua Bengio: Let’s take the major challenge of the current COVID-19 crisis. There are many candidates for the treatment of COVID-19, and they are being tested in a great number of clinical trials. But not every combination can be tested; there are simply too many. So the current trials might miss some synergistic effects: each drug given alone might show no big effect, but given in a certain combination they could be an effective treatment. These synergistic effects often occur at the level of cell pathways and mechanisms. For example, we already know about some synergistic effects in multidrug treatment of HIV. If we find out more about the underlying cellular mechanisms in COVID-19 and its treatment, we could identify synergistic effects that are possible using specific combinations of treatments. 

So it is not only about understanding what happens, but also about a kind of simulation that directs further research? 

Yoshua Bengio: Indeed, this is an important potential of AI in cellular dynamics: To create a model of complex processes. And to make predictions based on this model; to understand how to achieve an effect. 

Fabian Theis: Identifying and understanding the underlying mechanisms can already bring great benefits: In many cases, the knowledge generated inspires further research by scientists all over the world, e.g. to revert processes in the case of disease via drug treatments or other interventions. 

Yoshua Bengio: For example, cancer. If we understand these mechanisms better and in greater depth, we also have the potential to better understand the risk factors and therapeutic opportunities for many diseases – like diabetes, for example.

What role does big data play in all that? 

Yoshua Bengio: It is the basis of our research and for the application of AI. No human mind can absorb all these hundreds of thousands of data points – but a computer can!

Fabian, thanks to the Helmholtz network you have access to lots of data. The Helmholtz Zentrum München has one of the biggest biodata banks in Europe. You run the affiliated Helmholtz Institute of Computational Biology, which includes the Helmholtz International Lab “Causal Cell Dynamics”. What is the idea behind the Helmholtz International Labs? 

Fabian Theis: The selected research labs receive multi-annual funding from Helmholtz to expand international connections and networks. This creates a really productive and prolific environment, because we can afford to attract more promising PhD students and strengthen our international exchange and collaboration. As a result, we gain the essential ingredients for excellent research: funding and bright minds. The new collaboration with Mila brings benefits in both aspects for both institutions. We want to play this out together to make significant progress in both AI and basic biology research. 

Mila

Institute of Computational Biology

Helmholtz AI

Helmholtz International Labs

To form an International Lab, a Helmholtz Center join forces with a renowned complementary international partner institution, focusing on an innovative research topic that is of high strategic relevance for Helmholtz. The aim is to establish a visible research activity by Helmholtz in a location abroad and to create a lasting strategic institutional partnership as well. The International Lab “Causal Cell Dynamics” is being co-funded for five years by Helmholtz’s Initiative and Networking Fund.

29.04.2021 , Christian Heinrich
Print Version